A moving AI visibility curve raises one useful question, and it is not "what did the team do this month?" It is "what changed: the brand or the engine?" A visibility score measures a pair, your brand and the model answering. Replacing the model is enough to move the number, up or down, with no change to your site or campaigns and no alert firing.
For a marketing leader, the rule is simple: no movement gets reported, rewarded or funded until it has been dated against the model calendar. The check is cheap. Skipping it can cost a budget decision.
The GPT-6 rollout in ChatGPT, in October 2026, is the textbook case: a rollout that began over two days depending on the plan, sources that changed form, and measurement that kept running as if nothing had happened.
Can a model move your score when your brand has not changed?
Yes, because the model decides how long the answer is, how many sources it uses and whether it searches the web at all. All three weigh on your mentions and citations, and none depends on you.
Two sets of readings show it, pointing in opposite directions. Writesonic, which sells an AI visibility platform, ran the same 50 prompts through three ChatGPT models on 27 April 2026, once each, from one US account. The Nanga team separately observed, in October 2026, a move from one model to a newer one at another AI company.
| What moves | Before | After | Reading |
|---|---|---|---|
| Share of citations going to brand websites | 56.8% | 47.2% | Writesonic, GPT-5.4 to GPT-5.5 Thinking |
| Citations per answer | 9.4 | 7.2 | Writesonic, same models |
| Search queries per prompt | 10.5 | 7.3 | Writesonic, same models |
| Citations per answer | 3.9 | 6.3 | Nanga team, another AI company |
| Average answer length, in characters | 1,900 | 3,400 | Nanga team, same change |
Sources: Writesonic, April 2026; Nanga team, October 2026. Writesonic itself calls its numbers directional: one account, one run per prompt, one day.
The most telling gap sits inside a single product name. In Writesonic's reading, GPT-5.3 Instant sends only 13.4% of its citations to brand websites, against 56.8% for GPT-5.4 Thinking. "ChatGPT" is not one behavior. The Nanga team has a similar reading: two sibling models from one AI company searched the web on 3 prompts out of 10 and on 9 out of 10.
You cannot guess the direction in advance. A longer answer has more room to name a brand, a shorter one has less, and neither says anything about the brand. The rise is the trickier case, because nobody investigates good news.
What did the GPT-6 rollout change in October 2026?
The model and how answers are presented, with a start spread over two days by plan. According to TestingCatalog and 9to5Mac, OpenAI announced new GPT-6 models for everyday conversations on 7 October 2026, with a presentation layer called Intelligent UI that mixes text, charts, buttons and widgets depending on the question. GPT-6 Sol came first for Plus, Pro, Business and Enterprise, GPT-6 Luna the next day for Free and Go, which Borncity dates to 8 October. Borncity also stresses that 7 October marks the start of a gradual distribution, not its completion.
Two reading consequences follow. A launch date opens a transition, it does not flip a switch. And since those days, the word "ChatGPT" has covered at least two variants depending on the user's plan.
One point goes unnoticed in this coverage: none of the three describes how sources are now displayed. That is where the numbers move. The Nanga team observed that from 7 October, ChatGPT's sources left their usual place for small citation pills. A tracker still looking in the old place sees sourced answers almost vanish, without a single measurement erroring out, while ChatGPT never stopped citing.
Arithmetic illustration, fictional brand. Alba & Co is cited in 30 answers out of 100. If 40 answers in 100 lose their sources on reading, spread at random, the count falls to 18. That is 12 points gone, with no recommendation changed.
What is a series break?
A series break is a change in the engine, in the form of its answers or in the measurement that makes two points on one curve non-comparable, with no brand involvement.
Its defining trait: nothing breaks. The Nanga team counted seven format changes in two months in 2026, and none failed a run. Three examples show what this looks like on a dashboard.
- A source that moves. On Google AI Overviews, in October 2026, the Nanga team saw references move within the answer's format. A tracker that does not follow the move reads a sudden collapse in sourced answers, while Google never stopped citing.
- A refusal that reads as absence. A recent model can decline a prompt. The result is an answer naming no brand, and a curve counts it as an answer where you are missing.
- A truncated answer. On some sensitive topics, a content filter stops the answer partway through a web address, and the cut repeats identically. The topic appears to lose its sources when only the output was cut.
These cases pull a score down. Others inflate it, such as prompt echoes and interface text counted as mentions: see the piece on false AI visibility signals.
How do you tell a model change from ordinary noise?
Not from a single number, because answers vary widely even when nothing changes. Schulte, Bleeker and Kaufmann followed four sectors in a Swiss-German setting from 24 January to 20 March 2026: from one day to the next, source lists overlap by only 34 to 42%. Re-running the same prompts on the same day does no better, at 32 to 43%. Most of the instability is the randomness of answers, not the passage of time. Ahrefs found the same churn across more than 43,000 keywords in 2025: an AI Overview's text has a 70% chance of changing between observations and 45.5% of cited sources are new each time, while the meaning stays very close (0.95 cosine similarity).
So an isolated move proves nothing. A break shows in its shape, and you read that shape by crossing signals. The piece on how many prompts a reliable measure needs deals with size. This one deals with cause.
| What you look at | Brand effect | Model effect | Noise |
|---|---|---|---|
| Engines affected | Several, gradually | One, from a given date | No stable pattern |
| Competitors | You move relative to them | Everyone moves the same way | Scattered moves |
| Prompts | The ones you worked on | All topics | A few at random |
| Share of sourced answers | Stable | Clean jump | Oscillates around its level |
| Answer length and form | Stable | Change on the same date | Stable on average |
The author's reading grid, a starting hypothesis and not proof.
One criterion is not enough. Three lined up on the same date justify treating the move as a break until shown otherwise.
Which reading rules should you adopt?
Five, usable from your next review.
- Keep a model log. One line per event: date, engine, what changed, source. Note the start and likely end of a rollout, since rollouts come in stages. Without the log, every move gets credited to the latest marketing action.
- Watch the share of sourced answers. This is the share of answers using at least one source, across all brands. It barely depends on you, which makes it a good control variable: if it jumps, look at the engine before your content. Compare it with its own history, not an outside benchmark: the 57.8% of ChatGPT runs without a citation found by Schulte and coauthors hold for their corpus and period.
- Compare like with like. Same prompts, same engines, same country, same language. Adding an engine or a batch of prompts changes the score without changing the brand. Read engine by engine: an all-engine average dilutes a break on one.
- Read yourself next to competitors. On the same answers, a move shared with rival.fr and maison-lin.com is about the engine. A move that is yours alone is about you.
- Handle the break honestly. Keep the curve, mark the date, set a new baseline over two to four weeks (the rolling window Schulte and coauthors recommend), and calculate no change across the break. Faced with the model change mentioned above, the Nanga team left the break visible in its series and flagged it, rather than stitching the curve. A report stating "break on 7 October, new baseline on 4 November" survives a CFO's question better than a stitched curve.
What happens to the budget when a break lands mid-campaign?
Make no call on the period that crosses it. A content campaign launched in September and measured at the end of October carries the arrival of GPT-6 in its result this year, whatever that result is. Cutting because the score fell or doubling down because it rose would fund a decision made by the engine.
The fix exists and has to be set up before the break: keep control prompts you never work on, and compare the two groups. The full method is in the guide to proving GEO incrementality.
Arithmetic illustration, fictional brand. On the prompts it worked on, Alba & Co goes from 20% to 26%. On control prompts, from 20% to 25%. The model explains most of it: the gain attributable to the work is 1 point, not 6, and that point still needs confirming over the following weeks.
What still lacks proof?
As of 29 October 2026, research for this article found no independent study measuring GPT-6's effect on brand visibility. Published readings come from tool vendors, Writesonic's covers an earlier model on a single day, and the Nanga team's observations have not been audited externally. The same research found no validated threshold for separating a break from noise, and no measure of how long a rollout takes on a consumer engine. The grid above is a way of reasoning. It tells you what to check, without guaranteeing the cause.
What should a dashboard tell you?
It should let you rule out explanations that have nothing to do with your brand, before you read the score itself. Five things do that.
- Whether the last run is complete. A partial run shown as a normal one manufactures a drop.
- The difference between zero and not measured. An engine with no data for a period is not an engine where you are absent.
- The detail by engine. A break shows engine by engine, rarely in an average.
- Dates on the curve. Engines added and removed, rollouts, actions completed.
- The answers behind the number. A break is confirmed by reading ten answers before and ten after.
Across several engines and several months, this discipline is the first thing to slip. Nanga's GEO module builds it into the screens. On the Performance trends chart, markers flag the dates an AI model joined or left the project, and event flags let you date the rest, such as a public rollout. Performance by model breaks down AI Visibility, the GEO form of the AI Visibility Index, engine by engine, and a Market benchmark sits next to your source share, showing the share of AI answers that use at least one source. A partial run is flagged as such, with the number of prompts and the engines involved, and topic and competitor screens read "not measured" rather than a false 0%.
To see your visibility curve with its model changes, engine by engine: request a demo.
Frequently asked questions
What is a series break in AI visibility?
A change in the engine, in the form of its answers or in the measurement that makes two points on one curve non-comparable, with no brand involvement. Good practice is to date the break and read two segments, with no growth figure calculated across them.
Can a new model lower a brand's numbers for no reason?
Yes, and raise them too. Writesonic reported in April 2026 that the share of citations going to brand websites fell from 56.8% to 47.2% between GPT-5.4 and GPT-5.5 Thinking, on 50 prompts with one run each. Treat it as directional.
When did GPT-6 reach ChatGPT?
From 7 October 2026 for Plus, Pro, Business and Enterprise, and the next day for Free and Go, according to TestingCatalog and 9to5Mac. Borncity dates that next day to 8 October and notes that 7 October marks the start of a gradual distribution, not its completion.
Should you recalculate your history after a model change?
Generally no. A history stitched together by hand implies a continuity the data does not have. Keep the curve as it is, mark the date, set a new baseline over two to four weeks, and say so in the report.
Which indicator should you check first when a score moves?
The share of answers using at least one source, engine by engine. If it jumps alone on a given date, suspect the engine first. Levels do not transfer: Schulte and coauthors found 57.8% of ChatGPT runs without a citation, on a Swiss-German corpus.
Sources:
- TestingCatalog, OpenAI rolls out GPT-6 with Intelligent UI in ChatGPT, 7 October 2026. Report on OpenAI's announcement.
- 9to5Mac, OpenAI brings GPT-6 to ChatGPT and debuts Intelligent UI, 7 October 2026.
- Borncity, GPT-6: OpenAI startet den Rollout für ChatGPT-Nutzer am 7. Oktober, 20 October 2026. Written with AI assistance, according to the publisher.
- Writesonic, GPT-5.5 Cites Brand Sites 47% of the Time. GPT-5.4 Did 57%., 28 April 2026, updated 7 May 2026. 50 prompts tested on 27 April, one run per prompt, one US account.
- Schulte, Bleeker and Kaufmann, Don't Measure Once: Measuring Visibility in AI Search (GEO), 8 April 2026. Preprint, four sectors, Swiss-German data from 24 January to 20 March 2026.
- Ahrefs, AI Overviews Change Every 2 Days (But Never Change Their Mind), 11 November 2025. More than 43,000 keywords over one month.
- Nanga team, observations on its own data, through October 2026, not externally audited.